Adversarial Examples for Hamming Space Search.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning-to-hash methods are popular for large-scale approximate nearest neighbor search due to their representation learning capabilities.
• Deep learning models, including those for image classification, are vulnerable to adversarial examples, raising security concerns for deep retrieval systems.

Purpose of the Study:

• To investigate the robustness of modern deep hashing models against adversarial perturbations.
• To propose a novel method for generating adversarial examples specifically for Hamming space search.

Main Methods:

• Hash Adversary Generation (HAG) was developed to craft adversarial examples for Hamming space search.
• HAG generates imperceptibly perturbed queries that lead to semantically irrelevant nearest neighbors in targeted hashing models.
• The study also explores combining heterogeneous perturbations for black-box attacks.

Main Results:

• Extensive experiments demonstrate HAG's success in generating adversarial examples with minimal perturbations that mislead targeted hashing models.
• The transferability of these adversarial perturbations across various settings was verified.
• A simple yet effective method for constructing adversarial examples for black-box attacks was presented.

Conclusions:

• Modern deep hashing models are vulnerable to adversarial attacks.
• HAG provides a viable method for generating adversarial examples to test and improve the security of deep hashing systems.
• The findings highlight the need for robust deep hashing models against adversarial perturbations.